Lachmannian Archetype Reconstruction for Ancient Manuscript Corpora

Armin HoenenNAACL 2015

TL;DR

This paper introduces the first computational approach to automatically reconstruct the "archetype" (the earliest version) of ancient texts from manuscript copies, and proposes a novel evaluation method for these reconstructions. By comparing a philologically-inspired algorithm with a bio-informatics approach, they demonstrate that automated methods can achieve high accuracy, potentially accelerating the creation of scholarly editions.

What Problem Does This Paper Address?

Scholars working with ancient manuscripts often face a challenge: multiple copies of the same text have survived, but they differ due to copying errors, corrections, and other historical processes. This leads to uncertainty about the original wording and the relationships between manuscripts. The traditional philological approach aims to reconstruct a single, authoritative version of the text – the "archetype" – which is believed to be closest to the author's original. This is often done by creating a "stemma," a diagram showing the lineage of manuscripts. However, manually reconstructing the archetype and the stemma is a laborious and subjective process. This paper seeks to automate the reconstruction of the archetype text, drawing parallels with methods used in bioinformatics for reconstructing ancestral DNA sequences.

Methods

The authors propose two main methods for reconstructing the archetype text, both using a pre-computed stemma as input:

  • Philological Reconstruction (Majority-Vote Bottom-Up Algorithm): This method mimics traditional philological practice. For each position in the text, it looks at all the direct "descendant" manuscripts of a hypothetical lost manuscript (represented by a node in the stemma). It then selects the most frequent variant among these descendants as the reconstructed variant for that position. If there's a tie, all variants are kept for later disambiguation. This process is applied recursively from the leaves of the stemma up to the root.
  • Bio-informatic Reconstruction (Statistical Approach): This method leverages established algorithms from bioinformatics. The manuscript texts are converted into a "pseudo-DNA" sequence format. Then, a phylogenetic tree is generated using a maximum likelihood criterion. Ancestral sequences (representing the archetypes) are then inferred at the internal nodes and the root of this tree using probabilistic models (marginal and joint probabilities). The paper uses the PAML software for this.

A novel evaluation method is also introduced, comparing the automatically produced archetype (PA) against the known "gold standard" archetype from artificial benchmark datasets. This evaluation considers both the overall agreement across the entire text (Whole Text Evaluation - WTE) and the agreement specifically at positions where variation exists (Proportion of Correctly Disambiguated Variation - PVE).

Datasets

The paper utilizes two artificial manuscript traditions for evaluation:

  • Parzival: An English text tradition.
  • Notre Besoin: A French text tradition.

These datasets are described as "fully digitized sets of manuscripts which have been produced through manual copying in recent times whilst recording the true copy history/genealogical relationships." They are provided in a word-aligned tabular format, simplifying collation. The authors note that they excluded a third tradition, "Heinrichi," due to issues with interpreting Old Finnish data.

Key Insights and Ideas

  • Bio-informatics as a powerful tool for philology: The paper demonstrates that algorithms developed for DNA sequence reconstruction can be effectively adapted for reconstructing ancient texts, offering a more systematic and potentially less subjective approach.
  • Majority-vote is a strong heuristic: The philological majority-vote approach, despite its simplicity, performs remarkably well, suggesting that the most frequent variant in surviving manuscripts often reflects the original reading.
  • Archetype reconstruction is distinct from stemma construction: The authors emphasize that a good stemma doesn't guarantee a good archetype, and vice-versa. These are separate but related tasks that can inform each other.
  • Automated reconstruction can accelerate scholarly work: The success of these automated methods suggests they could significantly speed up the creation of critical editions, a core task in philology.
  • The "Lachmannian" approach is computationally feasible: The paper successfully implements and evaluates an automated version of the Lachmannian method, which emphasizes active reconstruction and emendation, rather than the Bédierian approach of relying on a single best manuscript.

Key Results

  • The philological reconstruction (majority-vote) outperformed the bio-informatic reconstruction (PAML) in terms of WTE and PVE scores for both the Parzival and Notre Besoin traditions.
  • Specifically, for Parzival, the majority-vote achieved a WTE of 0.96 and PVE of 0.88, while PAML achieved 0.91 and 0.73 respectively.
  • For Notre Besoin, majority-vote achieved 0.97 WTE and 0.94 PVE, compared to PAML's 0.95 and 0.9.
  • Both automated methods significantly outperformed randomly generated archetypes and even a "maximally wrong" archetype, indicating their reliability.
  • The ML trees generated by PAML had a relatively low Average Sign Distance (ASD) score when compared to the true stemmas, suggesting that while PAML is good at reconstruction, its tree inference might be less precise for these specific datasets.

Limitations and Future Work

  • Limited scope of variation: The current models primarily handle word substitutions. Future work should address other types of textual variation, such as word deletions, insertions, word separation errors, missing passages, marginal text, unreadable text, and oral variation/contamination.
  • Assumption of a pre-generated stemma: Both reconstruction methods rely on a pre-computed stemma. The paper suggests that the inverse process – generating stemmas from an archetype – could be a fruitful area for future research.
  • Textual data format: The current approach requires word-aligned texts. Extending this to handle unaligned texts would be beneficial.
  • "Contamination" in texts: While the paper notes that contamination (where a manuscript is influenced by multiple sources) is a problem for stemmatology, it suggests it might be less severe for archetype reconstruction at the word level. Further investigation into this is warranted.
  • Enriching bio-informatics: The authors propose that the unique characteristics of manuscript traditions (e.g., the higher probability of having both an original and its copy present) could enrich the field of ancestral sequence reconstruction in biology.

Summary updated

BibTeX
@inproceedings{hoenen2015lachmannian,
  author = {Armin Hoenen},
  title = {{Lachmannian Archetype Reconstruction for Ancient Manuscript Corpora}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
}
Figure